An SoC based CNN Accelerator of Tiny-YOLOv2 for Blind Assistive System on FPGAs

Abstract:

With the development of Deep Learning, especially Convolutional Neural Networks (CNNs), there have been tremendous advances in neural networks for object detection. These models can accurately detect and categorize objects in a wide range of complex scenarios and have been practically applied in people's daily lives. However, neural networks have a large number of parameters and operations. Therefore, many methods and hardware architectures have been proposed to accelerate neural networks to process data efficiently. This paper proposes an assistive system for blind people and implements it on a system-on-chip (SoC). The system consists of an ARM CPU for text-to-speech and a neural network accelerator module for object detection tasks based on Tiny You-Only-Look Once version 2(Tiny YOLOv2) to make it easier for blind people to walk alone in unfamiliar environments.

Artitechture:

The architecture of our proposed can be divided into five steps: data preprocessing, data sequence processing, Tiny YOLOv2 calculation, NMS, and Text-to-Speech. The proposed hardware architecture is implemented on the EGO-ZU19EG FPGA, which is a SoC architecture containing an ARM Cortex-A53 CPU and an UltraScale+MPSOC XCZU19EG. The proposed hardware accelerator is in Programmable Logic (PL), which contains data memory (BRAM), convolution unit (CONV U), quantization unit and max pooling unit, and communicates with the Processing System (PS) using AXI bus protocol, where the PS contains the CPU and external memory DDR4. When PS needs to accelerate the convolutional neural network, it will first preprocess the feature map and weight data and then store them in DDR4. Then, PL will read the data from DRAM through Xilinx's DMA IP. When all the data for computation are transferred to data memory and weight memory, the PS side will be idle, and the PL will start performing the required computation.

Fig 1¡GArchitecture block diagram of the proposed accelerator.

Fig 2:The result of object detection on EGO-ZU19EG

Made by ³¯«T¤Í